MR imaging has evolved significantly over the last decades, improving the quality of diagnosis and treatment with the information extracted from its data. MRI can generate data showing various types of tissues without the use of radiation.
However, in comparison to other imaging modalities, such as CT and ultrasound, it is more susceptible to various kinds of artefacts, especially patient motion [224].
The problem is that any type of movement, voluntary or not, is faster than the time required to collect enough data to create an image [224]. The issue is present in most available MRI sequences. Motion and other artefacts can have a negative influence on the quality of diagnosis [54, 107] as well as on the performance of image processing methods, e.g. segmentation, classification, registration, and texture analysis [8, 9, 84, 94, 200]. Commonly, these artefacts are divided into three groups [53, 54]:
• patient-related artefacts
• signal-processing artefacts
• hardware/machine-related artefacts
Patient-related artefacts. Motion is the most common artefact overall. It can be periodic or bulk (i.e. rigid). Periodic motion, e.g. due to blood flow, cardiac and respiratory motion, creates discreteghost artefactsalong the phase-encode direction.
“Ghosting” is a partial or complete copy of a structure or object. In contrast, bulk motion results in diffuse noise spread broadly along the phase-encode direction.
Further patient-related artefacts can occur when metal objects are present in the patient’s body. These can cause image distortion, high signal, and even signal loss.
[53, 54, 107, 224]
In the current work, no subjects had any metal objects inside their bodies. Periodic motion was not observed and is generally not an issue in knee MRIs. Bulk motion of the subjects was limited due to the knee coil used, but could not be completely avoided (Fig. B.1).
Figure B.1:Motion artefacts in knee MRIs
Signal-processing-related artefacts. These can further be separated into chemical shift, partial volume, wrap around, Gibbs phenomenon, and more, artefacts [53, 54].
Chemical shift artefacts mainly occur at the interfaces between water and fat due to the difference in resonant frequency of the protons of these two tissues. This will cause a shift in the mapping of fat and water pixels when the image is created.
The chemical shift is present in form of dark and bright bands, often left and right of the tissue boundaries. The effect of this artefact is intensified with increasing magnetic field strength of MR machines. In MR images of the knee, as the ones from the current study, the chemical shift can appear at the edges of bones, causing difference in cartilage thickness. [53, 54, 107]
Partial volume effects occur when the signal from different tissues in a voxel are averaged out and will thus cause a loss of spatial resolution. These effects can be mitigated by choosing smaller pixels/voxels and/or a smaller slice thickness. [53, 54]
Another artefact related to signal-processing is the(phase) wrap-aroundoraliasing artefact. It occurs when the size of an anatomical structure is larger than the chosen FOV. This will fold the parts outside the FOV to the opposite side of the image. One possible solution is to increase the FOV and another one to perform phase oversampling, which is often already available in the MR-scanner software.
[53, 54, 107].
Several of the knee MRIs of this work showed overlapping structures, but almost exclusively the ones acquired with 1.5T MR-scanners (Fig. B.2).
The last of the most common signal-processing artefacts is theGibbs phenomenonor truncation/ringing artefact. It can appear in form of alternating and evenly spaced dark and bright bands close to sharp high-contrast boundaries. [53, 54, 107]
This effect was seen in some of the knee images, but generally just in individual slices and not on the entire dataset (Fig. B.3).
Figure B.2:Wrap-around artefacts observed in knee MRIs
Figure B.3:Ringing artefacts
Hardware/machine-related artefacts. This third and final group of artefacts is related to the MR machine and its components. Inhomogeneities of the external, i.e.
static, magnetic fieldB0and the gradient fieldsB1cause intensity inhomogeneities in the MR images [8, 9, 54, 84, 94]. To understand how these field inhomogeneities affect the outcome of the images, a brief background will follow next.
The static magnetic field causes hydrogen protons in the body to align and precess around its field direction. The gradient coils on the other hand, can be switched on to temporarily create perpendicular and oscillating fields, which cause a frequency variation of the protons along the direction of the gradient. When switched off again, the protons gradually align back towards the static field. The excitation of the protons will emit radio frequencies which can be measured. The use of three types of gradient fields (x,y, andz) allows the spatial encoding of the MR signal and finally the generation of the images. Thus, any inhomogeneity in either the static or gradient fields, will not evenly excite the protons at the desired location. The resulting images will exhibit spatial and/or intensity distortions (Fig. B.4). [53, 54]
Figure B.4:Intensity distortions
Many of the artefacts described above can only be mitigated or avoided at the time of the image acquisition. The ones related to magnetic field inhomogeneities can be corrected using an approach calledBias Field Correction which was introduced as a pre-processing step for the data of this work (section 4.2). A popular algorithm for BFC is N4ITK [211] and was used in this work.
C Augmentation
The augmentation of the knee images of this work had to be refined in comparison to the initial approach describe in [148]. When translation and rotation is applied to the already cropped images from section 4.3, parts of the structures exit and new pixels enter the image frame (Fig. C.1 – bottom row). The new pixels are not known in the cropped MRIs and therefore filled with zeros which causes the loss of information. However, in the full size images this information is available in most cases. Therefore, augmentation was performedpriorto the actual cropping to recover “lost” anatomical structures (Fig. C.1 – top row).
Figure C.1:Comparison of augmenting images before cropping (top row) and after cropping (bottom row)